Volume 86, Issue 4, Pages 1067-1077 (May 2015) Dynamic Control of Response Criterion in Premotor Cortex during Perceptual Detection under Temporal Uncertainty  Federico Carnevale, Victor de Lafuente, Ranulfo Romo, Omri Barak, Néstor Parga  Neuron  Volume 86, Issue 4, Pages 1067-1077 (May 2015) DOI: 10.1016/j.neuron.2015.04.014 Copyright © 2015 Elsevier Inc. Terms and Conditions

Figure 1 Detection Task and Dynamical Response Criterion (A) Behavioral task represented by the vertical position of the mechanical probe during a trial. The stimulator indented the skin of one fingertip of the restrained hand (“probe down”), and the monkey reacted by placing its free hand on an immovable key (“hold key”). After a variable prestimulus period (1.5–3.5 s), a vibratory 0.5 s stimulus was presented on half of the trials. At the end of a fixed delay period (3 s), the stimulator moved up (“probe up”), instructing the monkey to make a response movement to one of two push buttons. The pressed button indicated whether or not the monkey felt the stimulus. (B) The variability in stimulus onset times and the fixed delay period defined a 2 s temporal window of possible stimulation. No stimulus was delivered before 1.5 s or after 3.5 s from the “hold key” event. The window of possible stimulation was not explicitly cued to the animal. (C) A possible mechanism to efficiently solve the task requires modulating the response criterion (the strength of sensory evidence required to produce a stimulus-present response) over time. Outside the possible stimulation window, the response criterion is high to avoid false positives. Within the window, the response criterion decreases to allow correct detections. (D) The mechanism described in (C) could be dynamically implemented by a separatrix in the neural space, dividing the basins of attraction of two attractors. The black trace is a trajectory of a correct rejection trial. The blue traces represent a hit (ending in the “yes” attractor) or a miss trial (ending in the “no” attractor). The distance from the current neural state to the separatrix at each point in time represents the response criterion. Neuron 2015 86, 1067-1077DOI: (10.1016/j.neuron.2015.04.014) Copyright © 2015 Elsevier Inc. Terms and Conditions

Figure 2 Detection of False Alarm Events by Template Matching (A) Firing rate of three simultaneously recorded neurons during a single FA trial (green traces) and average response of the same neurons during hit trials (blue traces). The shaded bar indicates the stimulation period. (B) We used a 1 s segment of the averaged activity during hit trials (left, blue trace) as a template (inset) to detect FA events in single FA trials. FA events were identified in single FA trials (middle and right, green traces) on the basis of the mean squared error between the single FA trial firing rate and the template. Red lines indicate the start of the template. (C) The average activity over FA trials realigned according to the times of detected events (green trace) matches the average over hit trials (blue trace) even outside of the period used as template. In contrast, the same method applied to CR trials produces a much weaker match. (D) Histogram of differences in the detected FA times from pairs of simultaneously recorded neurons. A significant fraction of FA trials was detected at the same time compared to CR trials (black bars) and chance level (black line, p < 0.001). Chance level was obtained by shuffling the trials to disrupt the correspondence between detected FA events in simultaneously recorded neurons. Error bars indicate 95% confidence intervals. np is the number of neural pairs, and nFA and nCR are the number of FA and CR trials, respectively. Neuron 2015 86, 1067-1077DOI: (10.1016/j.neuron.2015.04.014) Copyright © 2015 Elsevier Inc. Terms and Conditions

Figure 3 Probability of False Alarm over Time (A) Mean relative frequency of detected FA events over time during the time course of the trial. The probability of a FA event increases during the period of possible stimulation (within orange lines). Relative frequency was calculated as the portion of FA events detected in each time bin relative to the number of FA trials in which a FA event was detected at any time bin. The mean histogram was obtained by averaging across nFA = 947 FA trials distributed in nsess = 144 sessions. Error bars represent SEM. (B) Same as (A) for CR trials. Neuron 2015 86, 1067-1077DOI: (10.1016/j.neuron.2015.04.014) Copyright © 2015 Elsevier Inc. Terms and Conditions

Figure 4 Two-Dimensional Projection of the Population Dynamics Average neural trajectories during hit (blue), miss (red), and CR (black) trials projected onto two task-related axes (stimulus amplitude and stimulus detection). The trajectories are plotted from the beginning of the trial (green circles) to end of the delay period (orange circles). Stimulus-present conditions are plotted until 1.5 s and realigned at the stimulus onset time. Thick blue and red traces indicate the period of stimulation. The thick black line denotes the possible stimulation window (1.5 s to 3.5 s). Units are arbitrary. The inset is the N-dimensional Euclidean distance between the CR neural trajectory and the neural state at the stimulus offset time during the miss condition (end of the thick red trace), as an estimate of the distance to the separatrix over time. See Figure S4 for the same analysis performed for each subject separately. Neuron 2015 86, 1067-1077DOI: (10.1016/j.neuron.2015.04.014) Copyright © 2015 Elsevier Inc. Terms and Conditions

Figure 5 A Recurrent Network Model Learns to Solve the Task (A) Recurrent network model of rate units provided with a start cue input, a noisy sensory channel, and a decision output. The start cue indicates the beginning of a new trial. The stimulus is modeled as a pulse corrupted by noise. The decision is extracted from a linear combination of rates after the delay period. We trained the initially random network by changing the output connections. Because of the feedback loop, this effectively alters the recurrent dynamics of the network. (B) Target signal of the FORCE algorithm. The information provided during training was restricted to the behavioral outcome on each trial. Thus, no information about the probability of stimulation over time was given during training. (C) “Psychometric” function of the trained model obtained as the frequency of stimulus-present responses as a function of stimulus amplitude. Neuron 2015 86, 1067-1077DOI: (10.1016/j.neuron.2015.04.014) Copyright © 2015 Elsevier Inc. Terms and Conditions

Figure 6 The Network Infers the Window of Possible Stimulation (A) The response criterion, defined as the lowest stimulus amplitude that drives the network to a stimulus-present response, decreases during the period of possible stimulation (within orange lines). The response criterion was obtained by systematically probing the network with a bisection protocol at each time to find “borderline” stimulus amplitudes. Thin lines represent single realizations of this protocol. Thick line is the mean of n = 10 realizations. The response criterion was normalized with its maximum value during the trial. Inset shows the results of training networks with different possible stimulation windows. PSW is the center of the possible stimulation window used during training; min RC is the time in which the response criterion reaches its minimum value. (B) Mean relative frequency of detected FA events over time in the model obtained by the same template-matching algorithm used for the experimental data. The probability of producing a FA increases during the period of possible stimulation (within orange lines). Relative frequency is defined as in Figure 3. The mean histogram was obtained by averaging across sessions. Error bars represent SEM. nfa, number of FA trials. Neuron 2015 86, 1067-1077DOI: (10.1016/j.neuron.2015.04.014) Copyright © 2015 Elsevier Inc. Terms and Conditions

Figure 7 Neural Trajectories of the Recurrent Neural Network Model Neural trajectories during a hit (blue), a miss (red), and a CR (black) trial projected in the same axes as in Figure 4. The three trajectories overlap during the beginning of the trial. The stimulus is applied (in the hit and miss conditions) at the middle of the possible stimulation window (thick black line in CR). The hit trajectory evolves to the “yes” attractor, while the miss and CR trajectories end in the “no” attractor. The gray dots are points close to the separatrix, estimated as the states achieved during “borderline” stimuli. Inset shows the distance between the network state during CR trials and the separatrix. Note that distance is measured in the high-dimensional space and therefore cannot be inferred from the 2D plot. The fixed-points analysis of the trained network revealed a saddle point mediating the decision between the two stable fixed points. The green traces represent the trajectories starting near the saddle point following its unstable direction. For better visualization of this figure, the simulations were run without noise in the sensory inputs, but the effects do not change under noisy stimuli. Neuron 2015 86, 1067-1077DOI: (10.1016/j.neuron.2015.04.014) Copyright © 2015 Elsevier Inc. Terms and Conditions